FlipItRight: Stable Pose-Targeted Throw-Flip Across Diverse Objects

Technical Summary: FlipItRight – Stable Pose-Targeted Throw-Flip Across Diverse Objects

Problem Statement
Robotic throwing is a high-difficulty manipulation task where control is lost once the object leaves the gripper, making the release state (position, orientation, linear velocity, and angular velocity) the sole determinant of the landing trajectory. While existing methods have largely focused on achieving accurate landing positions, achieving a pose-targeted throw-flip—simultaneously controlling landing position and orientation—is significantly more challenging. This difficulty is compounded for high-degree-of-freedom (DoF) manipulators due to the tight coupling between end-effector motion and joint configurations, creating a high-dimensional, non-convex planning problem. Furthermore, prior systematic evaluations of pose-targeted throw-flip across objects with varying shapes, sizes, and masses have been scarce.

Methodology: FlipItRight Framework
The authors propose FlipItRight, a model-based, two-stage planning framework designed for stable planar throw-flip (motion confined to the $yz$-plane, rotation about the $x$-axis). The core innovation is treating the object release state as an explicit intermediate representation that bridges landing-pose requirements and robot motion generation.

1. Object-Level Release State Planner:

   • This module generates candidate release states ($\xi_{rel}$) that satisfy the desired landing pose.
   • It solves a constrained nonlinear optimization problem to minimize landing position and orientation errors while penalizing high-energy states (rotational/translational kinetic and potential energy) to promote stable landings.
   • The optimization utilizes a ballistic propagation operator (Algorithm 1) to simulate flight dynamics under assumptions of rigid grasp, decoupled load dynamics, negligible aerodynamics, and instantaneous release.
   • Candidates are filtered based on strict thresholds for landing position ($\pm 5$ mm) and orientation ($\pm 0.5^\circ$).

2. Robot-Level Swing Motion Planner:

   • This module evaluates the kinematic executability of the candidate release states and constructs a feasible swing motion.
   • Swing-Posture Selector: It solves for the initial joint configuration ($q_0$), release configuration ($q_f$), and swing duration ($T_{end}$). Crucially, it enforces a structured near-release motion design: the end-effector velocity is required to converge to a target constant value over a stabilization interval $[T_{stab}, T_{end}]$. This design aims to improve robustness against release-timing uncertainty by minimizing the impact of timing errors on the realized release velocity (since $\Delta v \approx a \Delta t$, reducing acceleration $a$ near release reduces error).
   • The planner solves a bounded nonlinear least-squares problem to satisfy release-state residuals, velocity-constancy constraints, joint acceleration limits, and swing direction/height requirements.
   • Swing-Trajectory Generator: Once a feasible posture is selected, a reference trajectory is generated using cubic Hermite splines for the acceleration phase and constant-velocity motion for the final phase. An offline optimization refines the joint-space trajectory to track this reference while respecting joint limits.

Key Contributions

1. Systematic Real-Robot Evaluation: The paper presents the first real-robot study to systematically evaluate pose-targeted throw-flip across objects of varying shapes (blocks, cylinders, prisms), sizes, and masses, and across multiple target poses.
2. Explicit Intermediate Representation: The introduction of the object release state as a decoupled intermediate representation allows for principled candidate filtering and adaptive selection of robot configurations, separating object-level landing feasibility from robot-level executability.
3. Structured Near-Release Motion: The framework imposes approximately constant end-effector velocities during the final swing phase. This design choice is shown to improve robustness to release-timing uncertainty without requiring learned models or iterative data collection.
4. Model-Based Deployment: The approach relies on analytical flight dynamics and trajectory optimization, enabling direct deployment on new objects and targets without environment-specific calibration or prior data collection.

Experimental Results
The framework was validated on a UR10e robot with an OnRobot RG6 gripper across 120 trials involving four distinct objects (two rectangular blocks, an elliptical cylinder, and a hexagonal prism) and three target distances each.

• Success Rate: The system achieved an overall 90% success rate (108/120 trials), defined as the object landing stably on the intended support face with the correct orientation.
• Performance Variance: Performance varied by object geometry. Block 2 and the hexagonal prism achieved high success rates (often 10/10), while Block 1 and the elliptical cylinder showed slightly lower rates. The authors attribute Block 1's lower performance to a less instantaneous release transition (likely due to required higher grasping forces) and smaller landing face stability.
• Ablation Studies:
  • Initial Pose Selection: Replacing the optimized initial configuration with a fixed, hand-designed posture resulted in higher tracking errors in simulation, confirming the value of adaptive pose selection.
  • Near-Release Velocity Structure: Removing the constant-velocity interval (setting $\alpha=1.0$ instead of $0.7$) significantly reduced success rates and increased landing errors, particularly for objects with less instantaneous release characteristics. This confirms that the structured motion design reduces sensitivity to timing uncertainty.

Significance and Claims
The paper claims that FlipItRight demonstrates that high-DoF manipulators can achieve precise, pose-targeted throw-flip across diverse physical objects without relying on learning-based methods or extensive data collection. By decomposing the problem and explicitly modeling the release state, the framework achieves competitive or superior position accuracy compared to prior work (which often used coarser success criteria) while adding orientation control. The authors emphasize that the structured near-release motion design is a critical factor in achieving robustness in real-world execution, bridging the gap between idealized planning and the physical realities of gripper actuation and release timing. The work establishes a baseline for stable, pose-targeted throwing that can be directly deployed on new objects with minimal setup.